nano

Serve nano inference locally or on Modal and exercise the generation endpoints. Use this skill when the user wants to run the server, generate or extend audio from a checkpoint, cover a hummed melody, deploy the inference API, or test /generate /extend /cover with text, lyrics, melody, or style conditioning.

Add MP3s to the nano training corpus and run them through the full data-prep pipeline (upload, prepare, tokenize, optional melody/auto-tag/transcribe/ structure/key-detect, phonemize, pack). Use this skill when the user wants to add training data, ingest songs, build or grow the corpus, prepare data for training, run tokenize/melody/pack/tag/transcribe/structure/phonemize, or asks how to get their MP3s into the model.

Evaluate a nano checkpoint and pick the right diagnostic for the question. Use this skill when the user wants to evaluate a checkpoint, score sample quality (CLAP / spectral / silence / onset), check overfitting (train vs val), decide whether lyric conditioning is worth the compute, sweep sampling parameters, check genre coverage, or sanity-check DAC codec fidelity.

Audit and evaluate nano's lyric data and lyric conditioning. Use this skill when the user wants to analyze the lyric dataset, check transcribe/phonemize coverage or backfill progress, measure the Whisper-hallucination rate, see the vocal-gender or word-count distribution, decide whether the lyric filter should run, or evaluate whether a checkpoint actually sings the words.

Audit the health of nano's training data before a run — corpus size vs the scale rails, per-stream conditioning coverage (tags / structure / keys / lyrics / phonemes / melody), song-duration distribution, and tag/genre spread. Use this skill when the user wants to check whether the data is ready to train, see how many songs have each conditioning stream, check melody or phoneme coverage, audit corpus size, or find what data prep is still missing. For the lyric corpus specifically use eval-lyrics; for a trained checkpoint use eval-checkpoint.

Evaluate the health of a LIVE nano training run from its Modal logs — loss trajectory vs the random floor, per-codebook convergence order, val-checkup trend and strikes, throughput stability, ETA, and infra noise triage. Use this skill when the user asks "how is training going", "is the run healthy", "evaluate the training run", "is it diverging", or wants an ETA to a step count. For evaluating a finished CHECKPOINT (sample quality, WER, overfitting) use eval-checkpoint instead; for launching/resuming runs use train-model.

Run the nano pytest suite correctly. Use this skill when the user wants to run the tests, run a specific test file, run or skip the slow benchmark tests, or is writing or debugging a test and needs the fixtures and conventions.

Launch, monitor, resume, and pull checkpoints for nano training — local pipeline-validation runs and Modal multi-GPU (B200:4) DDP full runs. Use this skill when the user wants to train, start or resume a training run, kick off Modal training, read or diagnose training logs, check tok/s or per-codebook loss, or extract and download a checkpoint from the volume.